Specifically, desktop RAM is slabs of silicon, placed into little packages, soldered onto circuit boards in DIMM form or similar, to be plugged into a motherboard slot for RAM.

The AI demand is for the silicon itself, using advanced packaging techniques to put them on the same package as the complex GPUs with very high bandwidth. So these same pieces of silicon are not even being put into DIMMs, so that if they fall out of use they’ll be pretty much intertwined with chips in form factors that a consumer can’t easily make use of.

There’s not really an easy way to bring that memory back into the consumer market, even after the AI bubble bursts.

Still a pretty limited palette, everyone wearing the same color shirts.

PNG tends to fail hard with textures. For example, my preferred theme in my chess app, which has some wood grain textures, generates huge screenshot file sizes (2MB), whereas the default might be less than 10% as large. Similarly, when I screenshot this image the file size jumps to 2MB for a 0.8 megapixel image.

Rendered textured scenes could easily overload the PNG compression algorithm to where they’re huge, and if Discord is historically associated with gaming, one can imagine certain video game screenshots blasting past that 40mb limit.

I think HEIC plays friendly for how they store live photos: a container that has both a still image and a video of the surrounding time context. HEIC for the still photo and HEVC for the video probably optimizes the hardware acceleration for fast, low power processing of both parts of the data, and allows for a higher quality extraction of an alternative still photo from a different part of the video.

And maybe they want to have more third party support in place before they set JXL as a default. All the power and space savings in the world on capture might not mean as much if the phone has to do the work of exporting a JPEG or HEIC for each time that file interfaces with an app or the browser or whatever.

JPEG XL has a mode for losslessly encoding any lossy JPEG into a smaller file size without any loss of quality. Wikipedia has some description of general approaches for losslessly encoding JPEG files further.

I don’t know if webp uses any of these tricks, but I don’t see why it would be hard to imagine that compression artifacts from a 30-year-old format can be encoded more efficiently today.

Google didn’t kill JPEG XL. It might have set browser support back some, but there’s still a place for JPEG XL to take over.

All the modern video-derived formats (webp, heif/heic, avif) tend to be optimized for screen resolutions. But for print photography (including just plain old regular photography that wants to keep the option open of maybe printing some of the images eventually), the higher resolutions and higher quality stretches the limits of where those codecs actually perform well (in terms of file sizes, perceived quality, computational power of coding or decoding).

JPEG XL knocks the other modern images out of the water at those print resolutions and color spaces and quality. It’s not just for photography, either: medical imaging, archiving, printing, etc., all use much higher resolutions that what is supported on any screen.

And perhaps most importantly for future support, the iPhone now supports taking images in JPEG XL. If that becomes a dominant format for photographic workflows, to replace stuff like DNG and other raw formats, browser support won’t hold back the format’s adoption.

And if you already have compression artifacts, what use is lossless?

To further reduce file size without further reducing quality.

There are probably billions of jpeg files out there in the world already encoded in lossy JPEG, with no corresponding higher quality version actually available (e.g., the camera that captures the image and immediately saves it as JPEG). We shouldn’t simply accept that those file sizes are going to forever be stuck, and can think through codecs that further compress the file size losslessly from there.

It was the Joint Picture Experts Group that invented it, so Google had no ownership over it, unlike WebP.

No, JPEG called for submission of proposals to define the new standard, and Google submitted its own PIK format, which provided much of the basis for what would become the JXL standard (the other primary contribution being Cloudinary’s FUIF).

Ultimately, I think most of the discussion around browser support thinks too small. Image formats are used for web display, sure, but they’re also used for so many other things. Digital imaging is used in medicine (where TIFF dominates), print, photography, video, etc.

I’m excited about JPEG XL as a replacement for TIFF and raw photography sensor data, including for printing and medical imaging. WebP, AVIF, HEIF, etc. really are only aiming for replacing web distributed images on a screen.

If you screenshot computer/phone interfaces (text, buttons, lots of flat colors with adjacent pixels the exact same color), the default PNG algorithm does a great job of keeping the file size small. If you screenshot a photograph, though, the PNG algorithm makes the file size huge, because it’s just really poorly optimized for re-encoding images that are already JPG.

AI drives 48% increase in Google emissions

That’s not even supported by the underlying study.

Google’s emissions went up 48% between 2019 and 2023, but a lot of things changed in 2020 generally, especially in video chat and cloud collaboration, dramatically expanding demand for data centers for storage and processing. Even without AI, we could have expected data center electricity use to go up dramatically between 2019 and 2023.

That’s kinda always been how technology changes jobs, though, by slowly making the job one of supervising the technology. I’m no longer carving a piece of wood myself, but I’m running the CNC machine by making sure it’s doing things properly and has everything it needs to work properly. I’m not physically stabbing the needle through the fabric every time, myself, but I am guiding the sewing machine path on that fabric. I’m not feeding fuel into the oven to maintain a particular temperature, but I am relying on the thermocouple to turn the heating element on and off to maintain the assigned equilibrium that I’ll use to bake food.

Many jobs are best done as a team effort between human and machine. Offloading the tedious tasks to the machine so that you can focus on the bigger picture is basically what technology is for. And as technology changes, we need to always be able to recalibrate which tasks are the tedious ones that machines do better, and which are the higher level decisions best left to humans.

‘what if we did the same thing, but twice?’

That’s just a perfect way to describe it.

Why does this image look like an AI-generated screenshot? The letter spacing and weights are all wrong.

It’s like the relationship between mathematics and accounting. Sure, almost everything accountants do involve math in some way, but it’s relatively simple math that is a tiny subset of what all of mathematics is about, and the actual study of math doesn’t really touch on the principles of accounting.

Computer science is a theoretical discipline that can be studied without computers. It’s about complexity theory and algorithms and data structures and the mathematical/logical foundations of computing. Actual practical programming work doesn’t really touch on that, although many people are aware of those concepts and might keep them in the back of their mind while coding.

People who get downvoted a lot end up with a ‘low reputation’ indicator next to their name. You’ll know it when you see it.

Upvotes in meme communities do not add to reputation.

I think any kind of reputation score should be community specific. There are users whose commenting style fits one community but not another, and their overall reputation should be understood in the context of which communities actually like them rather than some kind of global average.

“I don’t want it to be like the butt of today’s jokes, I want it to be like the butt of 30-year-old jokes”

MacBook seamless suspend/sleep performance is like 25% of why my personal daily driver is MacOS. Another 50% is battery life, of which their sleep/suspend management plays a part. I’ve played around with Linux on Apple hardware but it’s just never quite been there on power management or sleep/wake functionality. Which is mostly Apple’s fault for poor documentation and support for other OS’s, but it just is, and I got sick of fighting it.

The Walkman and other tape players were so much superior to CD players for portability and convenience. Batteries lasted a lot longer for portable tape players than for CD players. Tapes could be remixed easily so you could bring a specific playlist (or 2 or 3) with you. Tapes were much more resilient than CDs. The superior audio quality of CDs didn’t matter as much when you were using 1980’s era headphones. Or, even if you were using a boombox, the spinning of a disc was still susceptible to bumps or movement causing skips, and the higher speed motor and more complex audio processing drained batteries much faster. And back then, rechargeable batteries weren’t really a thing, so people were just burning through regular single use alkaline batteries.

It wasn’t until the 90’s that decent skip protection, a few generations of miniaturization and improved battery life, and improved headphones made portable CDs competitive with portable tapes.

At the same time, cars started to get CD players, but a typical person doesn’t buy a new car every year, so it took a few years for the overall number of cars to start having a decent number of CD players.

They’re actually only about 48% accurate, meaning that they’re more often wrong than right and you are 2% more likely to guess the right answer.

Wait what are the Bayesian priors? Are we assuming that the baseline is 50% true and 50% false? And what is its error rate in false positives versus false negatives? Because all these matter for determining after the fact how much probability to assign the test being right or wrong.

Put another way, imagine a stupid device that just says “true” literally every time. If I hook that device up to a person who never lies, then that machine is 100% accurate! If I hook that same device to a person who only lies 5% of the time, it’s still 95% accurate.

So what do you mean by 48% accurate? That’s not enough information to do anything with.

Wouldn’t a louder room raise the noise floor, too, so that any quieter signal couldn’t be extracted from the noisy background?

If we were to put a microphone and recording device in that room, could any amount of audio processing be able to extract the sound of the small server out from the background noise of all the bigger servers? Because if not, then that’s not just a auditory processing problem, but a genuine example of destruction of information.

Yeah, you’re describing an algorithm that incorporates data about the user’s previous likes. I’m saying that any decent user experience will include prioritization and weight of different posts, on a user by user basis, so the provider has no choice but to put together a ranking/recommendation algorithm that does more than simply sorts all available elements in chronological order.